1063-7397/01/3001- $25.00 © 2001 MAIK “Nauka /Interperiodica”
Russian Microelectronics, Vol. 30, No. 1, 2001, pp. 54–58. Translated from Mikroelektronika, Vol. 30, No. 1, 2001, pp. 63–67.
Original Russian Text Copyright © 2001by Luk’yanov.
In this work, we suggest a circuit-design neuron
model in which information is represented as the mean
of stochastic sequences (streams). According to [1–4],
such neurons will be called stream neurons.
Stream neurons make it possible to reduce hardware
costs, in particular, in implementing neural networks.
Using stream neurons, designers can develop large net-
works and increase the number of synapses. One more
advantage of stream neurons is their relatively high
immunity to data transfer errors.
In , a stream neuron in which the number of ran-
dom bit generators is proportional to the number of
synapses is considered. In , the number of these gen-
erators was reduced to , where
is the number
In our circuit, the number of random bit generators
is small and does not depend on the number of syn-
We also consider an exact model of a stream neuron.
For learning, an approximate linear model is used.
Results of learning with the approximate model are
checked by accurately computing the neuron function
and also with the use a program model.
2. BASIC ELEMENTS OF A NEURON
The neuron being considered has
inputs to which
) are applied (where
+ 1) synaptic coefficients
, and a
, where some aggregate value is accumulated.
formed according to the sign of the reg-
ister value and the coefﬁcient of rarity
is applied to
the output of the neuron.
are coded by
bits and the
sign bit, and the register
is coded by (
+ 2) bits
and is represented as a one’s complement. The coefﬁ-
is coded by
bits. We assume that the parame-
ters take the following values:
The neuron incorporates
(1) bit logic gates in amount of the order of
(2) two quasi-summation lines of capacity
(3) four adders of capacitors
' + 1,
' + 2,
random bit generators and about
The ﬁgure shows the circuit of the neuron. Blocks
contain registers with appropriate synaptic
coefﬁcients and a choice device. The block
, as well as several adders. The block
contains the register
and implements stochastic fil-
3. OPERATION OF THE NEURON
Introduce necessary designations. First, we will
deﬁne the function
, which takes the value of
th bit of a nonnegative number
Then, we put
Now we deﬁne the sign function :
A Circuit-Design Model of a Digital Neuron Operating
with the Mean Value of a Stochastic Stream
A. V. Luk’yanov
Demidov State University, ul. Sovetskaya 14, Yaroslavl, 150000 Russia
Received January 25, 2000
—A digital neuron circuit suggested in this work consists only of simple logic gates, several adders,
and several random number generators. This makes it possible to reduce hardware costs and increase the num-
ber of synapses. A mathematical model and experiments made with a model program are described. The Rprop
optimization algorithm is used for learning.